S1 Version numbers for R and all packages

## [1] "R version 4.5.1 (2025-06-13)"
## [1] "tidyverse version 2.0.0"
## [1] "knitr version 1.50"
## [1] "ggpubr version 0.6.0"
## [1] "ggrain version 0.0.4"
## [1] "Hmisc version 5.2.3"
## [1] "rstatix version 0.7.2"
## [1] "emmeans version 1.11.1"
## [1] "flextable version 0.9.9"
## [1] "officer version 0.6.10"
## [1] "english version 1.2.6"

S2 Inclusion and exclusion criteria

Inclusion criteria applying to all participants:

Additional inclusion criteria for ASD participants:

Additional inclusion criteria for BPD participants:

Additional inclusion criteria for comparison participants:

Furthermore, we applied the following exclusion criteria:

Testing was discontinued if participants withdrew their consent, exclusion criteria applied or in the case of technical difficulties.

S3 Measures

We collected the following self-report questionnaires:

For all participants recruited for this project, we also collected the short version of the Borderline Symptom List, BSL-237.

For 37 of the dyads, we used a plexiglass screen placed between the interaction partners on the table to decrease the chances of spreading an undetected infection. We applied a transparent anti-reflection foil to reduce any mirroring effects. During the conversation, participants took off their face masks. We asked these participants how much the plexiglass influenced them during the interactions (plexi; scale from 0 to 3). All participants were asked how much the cameras influenced their behaviour (video; scale from 0 to 3). Rapport is a sum of the following ratings, all on scales from 0 to 3, thus ranging from 0 to 15:

S3.1 Group comparisons based on diagnostic status

measurement ASD BPD COMP BPDvsASD COMPvsASD COMPvsBPD
ADC_total 49.06 (±17.17), n = 17 76.95 (±14.98), n = 21 34.99 (±23.86), n = 82 0.006* 0.243 0.000*
AQ_total 34.29 (±6.18), n = 17 23.67 (±4.72), n = 21 14.57 (±5.27), n = 82 0.000* 0.000* 0.000*
BDI_total 15.94 (±11.57), n = 17 24.81 (±10.48), n = 21 3.99 (±3.79), n = 82 0.715 0.000* 0.000*
BERT.acc 0.80 (±0.07), n = 17 0.82 (±0.09), n = 21 0.83 (±0.07), n = 82 1.000 0.919 1.000
BERT.rt 5.79 (±2.78), n = 17 3.84 (±1.24), n = 21 3.33 (±1.33), n = 82 0.453 0.000* 0.637
BSL_total NaN (±NA), n = 0 46.71 (±21.44), n = 21 7.19 (±6.84), n = 37 NA NA 0.000*
IQ.estimate 116.47 (±13.90), n = 17 109.60 (±9.54), n = 20 112.77 (±13.26), n = 82 1.000 1.000 1.000
SMS_total 5.47 (±2.81), n = 17 11.67 (±2.11), n = 21 9.80 (±2.62), n = 82 0.000* 0.000* 0.027*
SPF_total 37.00 (±7.20), n = 17 40.95 (±8.36), n = 21 44.73 (±5.82), n = 82 1.000 0.000* 0.328
TAS_total 61.76 (±11.03), n = 17 55.33 (±11.60), n = 21 37.55 (±8.53), n = 82 1.000 0.000* 0.000*
age 37.59 (±13.19), n = 17 28.38 (±10.02), n = 21 28.89 (±10.18), n = 82 0.506 0.243 1.000
plexi 0.82 (±0.53), n = 17 0.85 (±0.90), n = 13 0.86 (±0.60), n = 43 1.000 1.000 1.000
rapport 12.24 (±2.19), n = 17 11.95 (±2.64), n = 21 12.33 (±2.39), n = 81 1.000 1.000 1.000
video 0.71 (±0.69), n = 17 0.81 (±0.75), n = 21 0.63 (±0.68), n = 82 1.000 1.000 1.000 
##       ilabel
## gender ASD BPD COMP
##    fem   6  14   52
##    mal  11   7   30
## 
##  Pearson's Chi-squared test
## 
## data:  tb.gen
## X-squared = 5.1108, df = 2, p-value = 0.07766

S4 Features

S4.1 Facial expressions extracted from OpenFace

We only included data of participants with a mean confidence of tracked frames greater than 75% and more than 90% successfully tracked frames. Facial expressions were captured as action units. We did not extract emotional expressions from these facial expressions as coherence between facial expressions and emotions is not a given and might be even less so for autistic people8.

For the calculation of synchronisation, we included rotational parameters (yaw, roll, pitch) as well as the same action units as in our previous study9:

  • Mealplanning: 1, 2, 6, 7, 9, 14, 15, 17, 20, 25, 26 and 45
  • Hobbies: 1, 2, 6, 7, 9, 15, 17, 20, 23, 25, 26 and 45

We also extracted total facial expressiveness as mean intensity of all action units for each interaction partner to be included in the MovEx and the CROSSturn models. For the CROSSturn model, we also included other action units, as listed below.

These correspond to the following movements:

  • AU1: inner brow raiser
  • AU2: outer brow raiser
  • AU4: brow lowerer (only CROSSturn)
  • AU5: upper lid raiser (only CROSSturn)
  • AU6: cheek raiser
  • AU7: lid tightener
  • AU9: nose wrinkler
  • AU10: upper lip raiser (only CROSSturn)
  • AU12: lip corner puller (only CROSSturn)
  • AU14: dimpler
  • AU15: lip corner depressor
  • AU17: chin raiser
  • AU20: lip stretcher
  • AU23: lip tightener
  • AU25: lips part
  • AU26: jaw drop
  • AU45: blink

Furthermore, we used translational head position parameters to infer head motion using the following formula with \(\Delta_t\) referring to the respective frame-to-frame changes:

\[\text{head movement} = \sqrt{\Delta_t x^2 + \Delta_t y^2 + \Delta_t z^2}\]

S4.2 Motion quantity extracted using Motion Energy Analysis

This figure shows body (red and purple) and head (yellow and orange) regions of interests of each interaction partner separately:

There was always a space between the head and body region. Regions were chosen such that they cover the full range of motion throughout one conversation of one interaction partner. Thus, their sizes differed which is why we scaled all values.

In addition to using the motion quantity to compute synchronisation, we also extracted total movement in each region of interest for each interaction partner to be included in the MovEx model.

S4.3 Speech and turn-taking features

We extracted pitch using praat’s autocorrelation method, a technique widely recognized for its reliability and accuracy10. We implemented a two-step pitch extraction method, as outlined by Hirst11. First, to capture a broad range of frequencies, we set a low pitch floor of 50 Hz and a high pitch ceiling of 700 Hz, with a time step of 15 ms. All other parameters were set to Praat’s default values. Second, using these initial pitch values, we determined the first and third quartiles of pitch for each participant and task. We then used these quartiles to compute individual pitch floors and ceilings with the following algorithm:

\[\text{floor} = \min\left( 0.75 \cdot Q_{1, hobbies}, 0.75 \cdot Q_{1, mealplanning}\right)\]

\[\text{ceiling} = \max\left( 2.5 \cdot Q_{3, hobbies}, 2.5 \cdot Q_{3, mealplanning}\right)\]

We then used these individual pitch floors (range = 46 to 168Hz, mean = 109.7 ± 31.8) and ceilings (range = 250 to 839Hz, mean = 481.7 ± 147.4) to extract pitch. To ensure an equal number of frames for all participants, we maintained a consistent time step across all analyses. By default, praat calculates this time step using the following formula:

\[\text{timestep} = \frac{0.75}{\text{floor}}\]

Here, we used the same time step of 0.016 as in our previous study12 which was determined based on the minimum individual pitch floor of that sample. Since the new sample did not include anyone with a lower pitch floor, the time step fits both samples.

Intensity was extracted by convolving the squared sound with a Gaussian analysis window. We used praat’s default values of minimum pitch 100Hz and time step of 0.01s.

To estimate synchrony, we extracted continuous pitch and intensity time series for every millisecond of the recording. For pitch extraction, we used consistent parameters across all participants instead of individualized settings. This was necessary because the analysis width depends on the pitch floor. Given the heterogeneity of our sample, we opted for a wide range of considered frequencies, setting the pitch floor at 50Hz and the pitch ceiling at 700Hz. For intensity, we relied on Praat’s default values.

In the case of turn-based synchronization, we correlated the median pitch or intensity of each turn with the median pitch or intensity of the preceding turn.

Next, we used the uhm-o-meter13,14 to differentiate between periods of speaking and silence, identify syllables and extract several prosodic features (total number of syllables, total number of silent phases, duration of speaking as phonation time, speech rate as number of syllables per second, articulation rate as number of syllables per phonation time, average syllable duration and silence-to-turn ratio). The resulting speaking and silent instances were visually and aurally inspected to verify the accuracy of the algorithm.

S4.4 Cross-modal features

We captured two types of cross-modal features:

  • Interpersonal synchronisation of one person’s head with the other person’s body movement and vice versa
  • AU activation, body and head movement during listening and speaking

S4.5 Synchrony computations

We used the following settings for our windowed lagged cross-correlation (WLCC):

WLCC settings in seconds
measure window step lag
Facial action units synchronisation 7 4 2
Body MEA synchronisation 30 15 5
Head MEA synchronisation 30 15 5
Intrapersonal synchrony 30 15 5
Pitch synchrony 16 8 2
Intensity synchrony 16 8 2

For each window, the maximum correlation value was chosen out of all relevant lags (peack-picking). We cross-correlated head movements (from OpenFace) with body motion energy time series (from MEA) to estimate intrapersonal synchrony.

S4.6 Feature lists

This list shows all features without the information of conversation, i.e., each of these features was added twice to the model, once from the mealplanning and once from the hobbies conversation. The number of features per model is displayed as well. For many of the extracted features, we calculated summary scores some which are indicated by abbreviations (mean, md = median, sd = standard deviation, min = minimum, max = maximum, ske = skewness and kurtosis)

model features no of features
BODYsync min_M_bodysync, max_M_bodysync, sd_M_bodysync, mean_M_bodysync, md_M_bodysync, skew_M_bodysync, kurtosis_M_bodysync 14
CROSSsync min_M_LOF, min_M_ROF, max_M_LOF, max_M_ROF, md_M_LOF, md_M_ROF, mean_M_LOF, mean_M_ROF, sd_M_LOF, sd_M_ROF, kurtosis_M_LOF, kurtosis_M_ROF, skew_M_LOF, skew_M_ROF 28
CROSSturn self_min_M_AU01_r, self_min_M_AU02_r, self_min_M_AU04_r, self_min_M_AU05_r, self_min_M_AU06_r, self_min_M_AU07_r, self_min_M_AU09_r, self_min_M_AU10_r, self_min_M_AU12_r, self_min_M_AU14_r, self_min_M_AU15_r, self_min_M_AU17_r, self_min_M_AU20_r, self_min_M_AU23_r, self_min_M_AU25_r, self_min_M_AU26_r, self_min_M_AU45_r, self_min_M_MEA_body, self_min_M_MEA_head, self_max_M_AU01_r, self_max_M_AU02_r, self_max_M_AU04_r, self_max_M_AU05_r, self_max_M_AU06_r, self_max_M_AU07_r, self_max_M_AU09_r, self_max_M_AU10_r, self_max_M_AU12_r, self_max_M_AU14_r, self_max_M_AU15_r, self_max_M_AU17_r, self_max_M_AU20_r, self_max_M_AU23_r, self_max_M_AU25_r, self_max_M_AU26_r, self_max_M_AU45_r, self_max_M_MEA_body, self_max_M_MEA_head, self_md_M_AU01_r, self_md_M_AU02_r, self_md_M_AU04_r, self_md_M_AU05_r, self_md_M_AU06_r, self_md_M_AU07_r, self_md_M_AU09_r, self_md_M_AU10_r, self_md_M_AU12_r, self_md_M_AU14_r, self_md_M_AU15_r, self_md_M_AU17_r, self_md_M_AU20_r, self_md_M_AU23_r, self_md_M_AU25_r, self_md_M_AU26_r, self_md_M_AU45_r, self_md_M_MEA_body, self_md_M_MEA_head, self_mean_M_AU01_r, self_mean_M_AU02_r, self_mean_M_AU04_r, self_mean_M_AU05_r, self_mean_M_AU06_r, self_mean_M_AU07_r, self_mean_M_AU09_r, self_mean_M_AU10_r, self_mean_M_AU12_r, self_mean_M_AU14_r, self_mean_M_AU15_r, self_mean_M_AU17_r, self_mean_M_AU20_r, self_mean_M_AU23_r, self_mean_M_AU25_r, self_mean_M_AU26_r, self_mean_M_AU45_r, self_mean_M_MEA_body, self_mean_M_MEA_head, self_sd_M_AU01_r, self_sd_M_AU02_r, self_sd_M_AU04_r, self_sd_M_AU05_r, self_sd_M_AU06_r, self_sd_M_AU07_r, self_sd_M_AU09_r, self_sd_M_AU10_r, self_sd_M_AU12_r, self_sd_M_AU14_r, self_sd_M_AU15_r, self_sd_M_AU17_r, self_sd_M_AU20_r, self_sd_M_AU23_r, self_sd_M_AU25_r, self_sd_M_AU26_r, self_sd_M_AU45_r, self_sd_M_MEA_body, self_sd_M_MEA_head, self_kurtosis_M_AU01_r, self_kurtosis_M_AU02_r, self_kurtosis_M_AU04_r, self_kurtosis_M_AU05_r, self_kurtosis_M_AU06_r, self_kurtosis_M_AU07_r, self_kurtosis_M_AU09_r, self_kurtosis_M_AU10_r, self_kurtosis_M_AU12_r, self_kurtosis_M_AU14_r, self_kurtosis_M_AU15_r, self_kurtosis_M_AU17_r, self_kurtosis_M_AU20_r, self_kurtosis_M_AU23_r, self_kurtosis_M_AU25_r, self_kurtosis_M_AU26_r, self_kurtosis_M_AU45_r, self_kurtosis_M_MEA_body, self_kurtosis_M_MEA_head, self_skew_M_AU01_r, self_skew_M_AU02_r, self_skew_M_AU04_r, self_skew_M_AU05_r, self_skew_M_AU06_r, self_skew_M_AU07_r, self_skew_M_AU09_r, self_skew_M_AU10_r, self_skew_M_AU12_r, self_skew_M_AU14_r, self_skew_M_AU15_r, self_skew_M_AU17_r, self_skew_M_AU20_r, self_skew_M_AU23_r, self_skew_M_AU25_r, self_skew_M_AU26_r, self_skew_M_AU45_r, self_skew_M_MEA_body, self_skew_M_MEA_head, other_min_M_AU01_r, other_min_M_AU02_r, other_min_M_AU04_r, other_min_M_AU05_r, other_min_M_AU06_r, other_min_M_AU07_r, other_min_M_AU09_r, other_min_M_AU10_r, other_min_M_AU12_r, other_min_M_AU14_r, other_min_M_AU15_r, other_min_M_AU17_r, other_min_M_AU20_r, other_min_M_AU23_r, other_min_M_AU25_r, other_min_M_AU26_r, other_min_M_AU45_r, other_min_M_MEA_body, other_min_M_MEA_head, other_max_M_AU01_r, other_max_M_AU02_r, other_max_M_AU04_r, other_max_M_AU05_r, other_max_M_AU06_r, other_max_M_AU07_r, other_max_M_AU09_r, other_max_M_AU10_r, other_max_M_AU12_r, other_max_M_AU14_r, other_max_M_AU15_r, other_max_M_AU17_r, other_max_M_AU20_r, other_max_M_AU23_r, other_max_M_AU25_r, other_max_M_AU26_r, other_max_M_AU45_r, other_max_M_MEA_body, other_max_M_MEA_head, other_md_M_AU01_r, other_md_M_AU02_r, other_md_M_AU04_r, other_md_M_AU05_r, other_md_M_AU06_r, other_md_M_AU07_r, other_md_M_AU09_r, other_md_M_AU10_r, other_md_M_AU12_r, other_md_M_AU14_r, other_md_M_AU15_r, other_md_M_AU17_r, other_md_M_AU20_r, other_md_M_AU23_r, other_md_M_AU25_r, other_md_M_AU26_r, other_md_M_AU45_r, other_md_M_MEA_body, other_md_M_MEA_head, other_mean_M_AU01_r, other_mean_M_AU02_r, other_mean_M_AU04_r, other_mean_M_AU05_r, other_mean_M_AU06_r, other_mean_M_AU07_r, other_mean_M_AU09_r, other_mean_M_AU10_r, other_mean_M_AU12_r, other_mean_M_AU14_r, other_mean_M_AU15_r, other_mean_M_AU17_r, other_mean_M_AU20_r, other_mean_M_AU23_r, other_mean_M_AU25_r, other_mean_M_AU26_r, other_mean_M_AU45_r, other_mean_M_MEA_body, other_mean_M_MEA_head, other_sd_M_AU01_r, other_sd_M_AU02_r, other_sd_M_AU04_r, other_sd_M_AU05_r, other_sd_M_AU06_r, other_sd_M_AU07_r, other_sd_M_AU09_r, other_sd_M_AU10_r, other_sd_M_AU12_r, other_sd_M_AU14_r, other_sd_M_AU15_r, other_sd_M_AU17_r, other_sd_M_AU20_r, other_sd_M_AU23_r, other_sd_M_AU25_r, other_sd_M_AU26_r, other_sd_M_AU45_r, other_sd_M_MEA_body, other_sd_M_MEA_head, other_kurtosis_M_AU01_r, other_kurtosis_M_AU02_r, other_kurtosis_M_AU04_r, other_kurtosis_M_AU05_r, other_kurtosis_M_AU06_r, other_kurtosis_M_AU07_r, other_kurtosis_M_AU09_r, other_kurtosis_M_AU10_r, other_kurtosis_M_AU12_r, other_kurtosis_M_AU14_r, other_kurtosis_M_AU15_r, other_kurtosis_M_AU17_r, other_kurtosis_M_AU20_r, other_kurtosis_M_AU23_r, other_kurtosis_M_AU25_r, other_kurtosis_M_AU26_r, other_kurtosis_M_AU45_r, other_kurtosis_M_MEA_body, other_kurtosis_M_MEA_head, other_skew_M_AU01_r, other_skew_M_AU02_r, other_skew_M_AU04_r, other_skew_M_AU05_r, other_skew_M_AU06_r, other_skew_M_AU07_r, other_skew_M_AU09_r, other_skew_M_AU10_r, other_skew_M_AU12_r, other_skew_M_AU14_r, other_skew_M_AU15_r, other_skew_M_AU17_r, other_skew_M_AU20_r, other_skew_M_AU23_r, other_skew_M_AU25_r, other_skew_M_AU26_r, other_skew_M_AU45_r, other_skew_M_MEA_body, other_skew_M_MEA_head 532
FACEsync min_M_AU01_r, max_M_AU01_r, sd_M_AU01_r, mean_M_AU01_r, md_M_AU01_r, skew_M_AU01_r, kurtosis_M_AU01_r, min_M_AU02_r, max_M_AU02_r, sd_M_AU02_r, mean_M_AU02_r, md_M_AU02_r, skew_M_AU02_r, kurtosis_M_AU02_r, min_M_AU06_r, max_M_AU06_r, sd_M_AU06_r, mean_M_AU06_r, md_M_AU06_r, skew_M_AU06_r, kurtosis_M_AU06_r, min_M_AU07_r, max_M_AU07_r, sd_M_AU07_r, mean_M_AU07_r, md_M_AU07_r, skew_M_AU07_r, kurtosis_M_AU07_r, min_M_AU09_r, max_M_AU09_r, sd_M_AU09_r, mean_M_AU09_r, md_M_AU09_r, skew_M_AU09_r, kurtosis_M_AU09_r, min_M_AU14_r, max_M_AU14_r, sd_M_AU14_r, mean_M_AU14_r, md_M_AU14_r, skew_M_AU14_r, kurtosis_M_AU14_r, min_M_AU15_r, max_M_AU15_r, sd_M_AU15_r, mean_M_AU15_r, md_M_AU15_r, skew_M_AU15_r, kurtosis_M_AU15_r, min_M_AU17_r, max_M_AU17_r, sd_M_AU17_r, mean_M_AU17_r, md_M_AU17_r, skew_M_AU17_r, kurtosis_M_AU17_r, min_M_AU20_r, max_M_AU20_r, sd_M_AU20_r, mean_M_AU20_r, md_M_AU20_r, skew_M_AU20_r, kurtosis_M_AU20_r, min_M_AU25_r, max_M_AU25_r, sd_M_AU25_r, mean_M_AU25_r, md_M_AU25_r, skew_M_AU25_r, kurtosis_M_AU25_r, min_M_AU26_r, max_M_AU26_r, sd_M_AU26_r, mean_M_AU26_r, md_M_AU26_r, skew_M_AU26_r, kurtosis_M_AU26_r, min_M_AU45_r, max_M_AU45_r, sd_M_AU45_r, mean_M_AU45_r, md_M_AU45_r, skew_M_AU45_r, kurtosis_M_AU45_r 168
HEADsync min_M_headsync, max_M_headsync, sd_M_headsync, mean_M_headsync, md_M_headsync, skew_M_headsync, kurtosis_M_headsync, min_M_pose_Rxsync, max_M_pose_Rxsync, sd_M_pose_Rxsync, mean_M_pose_Rxsync, md_M_pose_Rxsync, skew_M_pose_Rxsync, kurtosis_M_pose_Rxsync, min_M_pose_Rysync, max_M_pose_Rysync, sd_M_pose_Rysync, mean_M_pose_Rysync, md_M_pose_Rysync, skew_M_pose_Rysync, kurtosis_M_pose_Rysync, min_M_pose_Rzsync, max_M_pose_Rzsync, sd_M_pose_Rzsync, mean_M_pose_Rzsync, md_M_pose_Rzsync, skew_M_pose_Rzsync, kurtosis_M_pose_Rzsync 56
INTRAsync min_M_intra, max_M_intra, sd_M_intra, mean_M_intra, md_M_intra, skew_M_intra, kurtosis_M_intra 14
MovEx M_body_total_movement, M_head_total_movement, mean_intensity_M 6
Speech dyad_pit_sync_MEA_M_speech, dyad_int_sync_MEA_M_speech, dyad_spr_M_speech, dyad_str_M_speech, dyad_ttg_M_speech, dyad_no_turns_M_speech, nsyll_M_speech, npause_M_speech, pho_M_speech, art_M_speech, pit_sync_M_speech, int_sync_M_speech, art_sync_M_speech, pit_var_M_speech, int_var_M_speech 30

S5 Model performance

S5.1 Distinguishing BPD-involved from non-clinical interactions

While developing an algorithm for technology-assisted diagnostics of BPD was not the explicit goal of this research project, we explored the application of our features to the classification between BPD-involved and non-clinical interactions. Despite the features being chosen with symptoms and characteristics of ASD in mind, the CROSSturn, FACEsync, HEADsync and Speech models performed above chance in this comparison (BODYsync: pFDR = 1; CROSSsync: pFDR = 0.282; INTRAsync: pFDR = 1; MovEx: pFDR = 0.242). Specifically, the HEADsync model achieved 68.7% balanced accuracy (71.4; 65.9% specificity), the FACEsync 64% (64.3; 65.9% specificity), the Speech 61.6% (59.5; 63.6% specificity) and the CROSSturn 55.7% (52.4; 59.1% specificity). The stacking model performed comparable to the HEADsync and the MovEx model but outperformed the other base models (see [!T]), reaching 65.3% balanced accuracy (73.8; 56.8% specificity). Thus, the stacking model only misclassified eleven BPD-involved interactions as non-clinical, but 19 non-clinical interactions were labelled as BPD-involved.

S5.2 One-verus-One comparisons

comparison model sens spec BAC AUC p.fdr sig
ASD-COMP vs BPD-COMP BODYsync 32.353 47.619 39.986 0.361 1.000
ASD-COMP vs BPD-COMP CROSSsync 52.941 69.048 60.994 0.678 0.136
ASD-COMP vs BPD-COMP CROSSturn 67.647 85.714 76.681 0.749 0.000 *
ASD-COMP vs BPD-COMP FACEsync 76.471 59.524 67.997 0.719 0.000 *
ASD-COMP vs BPD-COMP HEADsync 55.882 57.143 56.513 0.609 0.936
ASD-COMP vs BPD-COMP INTRAsync 38.235 54.762 46.499 0.455 1.000
ASD-COMP vs BPD-COMP MovEx 58.824 78.571 68.698 0.758 0.027 *
ASD-COMP vs BPD-COMP Speech 64.706 76.190 70.448 0.771 0.000 *
ASD-COMP vs BPD-COMP STACK 70.588 92.857 81.723 0.805 NaN NA
ASD-COMP vs COMP-COMP BODYsync 58.824 61.364 60.094 0.607 0.000 *
ASD-COMP vs COMP-COMP CROSSsync 67.647 75.000 71.324 0.763 0.000 *
ASD-COMP vs COMP-COMP CROSSturn 44.118 72.727 58.422 0.627 0.000 *
ASD-COMP vs COMP-COMP FACEsync 79.412 70.454 74.933 0.780 0.000 *
ASD-COMP vs COMP-COMP HEADsync 58.824 68.182 63.503 0.608 1.000
ASD-COMP vs COMP-COMP INTRAsync 26.471 40.909 33.690 0.297 1.000
ASD-COMP vs COMP-COMP MovEx 64.706 72.727 68.717 0.777 0.000 *
ASD-COMP vs COMP-COMP Speech 64.706 72.727 68.717 0.673 0.000 *
ASD-COMP vs COMP-COMP STACK 79.412 81.818 80.615 0.845 NaN NA
BPD-COMP vs COMP-COMP BODYsync 28.571 36.364 32.468 0.308 1.000
BPD-COMP vs COMP-COMP CROSSsync 50.000 61.364 55.682 0.594 0.282
BPD-COMP vs COMP-COMP CROSSturn 52.381 59.091 55.736 0.569 0.045 *
BPD-COMP vs COMP-COMP FACEsync 64.286 63.636 63.961 0.670 0.007 *
BPD-COMP vs COMP-COMP HEADsync 71.429 65.909 68.669 0.736 0.000 *
BPD-COMP vs COMP-COMP INTRAsync 57.143 54.545 55.844 0.516 1.000
BPD-COMP vs COMP-COMP MovEx 69.048 70.454 69.751 0.686 0.242
BPD-COMP vs COMP-COMP Speech 59.524 63.636 61.580 0.693 0.007 *
BPD-COMP vs COMP-COMP STACK 73.810 56.818 65.314 0.760 NaN NA

S5.3 Multi-group comparisons

comparison model BAC p.fdr sig
MultiGroup BODYsync 48.6 1.000
MultiGroup CROSSsync 58.1 0.000 *
MultiGroup CROSSturn 57.3 0.000 *
MultiGroup FACEsync 57.6 0.007 *
MultiGroup HEADsync 54.6 0.000 *
MultiGroup INTRAsync 49.5 1.000
MultiGroup MovEx 60.8 0.000 *
MultiGroup Speech 63.9 0.000 *
MultiGroup STACK 63.2 NaN NA

S5.4 Gender comparisons

Did the models perform better for one gender than the other? We perform unpaired Wilcoxon tests for the labels and models separately.

S6 Model visualisations

The following figures were created inspired by NeuroMiner15 visualisations and show the sign-based consistency16 as well as the feature weights of the models distinguishing between interaction partners from ASD- and BPD-involved interactions. For models with a large number of features, the top 16 features are plotted.

S6.1 ASD-involved versus BPD-involved classifiers

S6.2 ASD-involved versus NC classifiers

S6.3 ASD-involved versus NC classifiers

S7 References

1.
Baron-Cohen, S., Wheelwright, S., Skinner, R., Martin, J. & Clubley, E. The Autism-Spectrum Quotient (AQ): Evidence from Asperger Syndrome/High-Functioning Autism, Males and Females, Scientists and Mathematicians. Journal of Autism and Developmental Disorders 31, 5–17 (2001).
2.
Paulus, C. Der Saarbrücker Persönlichkeitsfragebogen SPF (IRI) zur Messung von Empathie: Psychometrische Evaluation der deutschen Version des Interpersonal Reactivity Index. (2009).
3.
Popp, K. et al. Faktorstruktur and Reliabilität der Toronto-Alexithymie-Skala (TAS-20) in der deutschen Bevölkerung. Psychotherapie· Psychosomatik· Medizinische Psychologie 58, 208–214 (2008).
4.
Hautzinger, M., Bailer, M., Worall, H. & Keller, F. Beck-depressions-inventar (BDI). (1994).
5.
Graf, A. A German version of the Self-Monitoring Scale. Zeitschrift fur Arbeits- und Organisationspsychologie 48, 109–121 (2004).
6.
Kirby, A., Edwards, L., Sugden, D. & Rosenblum, S. The development and standardization of the Adult Developmental Co-ordination Disorders/Dyspraxia Checklist (ADC). Research in Developmental Disabilities 31, 131–139 (2010).
7.
Bohus, M. et al. The Short Version of the Borderline Symptom List (BSL-23): Development and Initial Data on Psychometric Properties. Psychopathology 42, 32–39 (2008).
8.
Costa, A. P., Steffgen, G. & Samson, A. C. Expressive Incoherence and Alexithymia in Autism Spectrum Disorder. J Autism Dev Disord 47, 1659–1672 (2017).
9.
Koehler, J. C. et al. Machine learning classification of autism spectrum disorder based on reciprocity in naturalistic social interactions. Transl Psychiatry 14, 1–9 (2024).
10.
Boersma, P. & Weenink, D. Praat: Doing phonetics by computer. (2022).
11.
Hirst, D. The analysis by synthesis of speech melody: From data to models. Journal of Speech Sciences 1, 55–83 (2011).
12.
Plank, I. S., Koehler, J. C., Nelson, A., Koutsouleris, N. & Falter-Wagner, C. Automated extraction of speech and turn-taking parameters in autism allows for diagnostic classification using a multivariable prediction model. Frontiers in Psychiatry 14, 1257569 (2023).
13.
De Jong, N. H., Pacilly, J. & Heeren, W. Uhm-o-meter [Computer software]. (2021).
14.
De Jong, N. H., Pacilly, J. & Heeren, W. PRAAT scripts to measure speed fluency and breakdown fluency in speech automatically. Assessment in Education: Principles, Policy and Practice 28, 456–476 (2021).
15.
Koutsouleris, N., Vetter, C., Wiegand, A., Hahn, L. & Mena, S. Neurominer. (2024).
16.
Gómez-Verdejo, V., Parrado-Hernández, E., Tohka, J. & Initiative, A. D. N. Sign-consistency based variable importance for machine learning in brain imaging. Neuroinformatics 17, 593–609 (2019).